Consider the script run_regressor_evaluation.py.
Apart from empty lines establishing semantic blocks, the script is unstructured and could be categorised as (a not very severe form of) "spaghetti code".
There is only one function main, which does everything, i.e.
- it loads a dataset,
- it projects and scales the data,
- it splits the dataset, and
- it creates and evaluates four types of models.
Notice that ...
- readability is mediocre, because the code is somewhat lengthy and low-level;
- there is some repetition; we are essentially repeating the same code for all four models;
- if we wanted to modify the evaluation (e.g. by changing the metric), we would need to make the changes in all four pieces of code pertaining to evaluations;
- there is no parametrisation; all evaluation parameters are hard-coded;
- we cannot easily reuse the evaluation code (other than by copying it);
- the data pipeline is hard-coded and used for all models simultaneously; there is no simple way of adapting the pre-processing exclusively for one particular model.
We shall address the first five points in the following. (For point (6), please refer to our refactoring journey, which will address it in detail.)